The failure is in the routing. An access request and a deletion request come through the same inbox, get read by the same agent, and land in a ticket taxonomy where privacy is one bucket with one macro attached. Crunchbase told WIRED that its misclassified reply came from a person on its customer success team rather than a generative AI tool [8]. Crunchbase offered that as reassurance. It describes the harder version of the problem: a human read an email that said "I am not requesting deletion at this time" [6] and permanently deleted the account anyway [7].
Timing says the same thing. The request went out on August 17 and the deletion confirmation came back two days later [6][7]. CCPA gives companies 45 days to complete a request [3], which put the real deadline at October 1 [16]. The wrong answer cleared the queue after roughly 4 percent of the available window [15], so no lateness report was ever going to surface it.
BeenVerified failed differently. The reporter wrote to its dedicated CCPA compliance address on the morning of August 19 [9] and got back a note saying his person report had already been removed from Person Search results, along with the phone number and email address he had supplied [10]. When he clarified that he had asked for access, the reply 15 minutes later denied the claim and said the company could not verify his identity, in a thread where it had already located his details [11]. WIRED disclosed that the reporter used generative AI to draft his bureaucratic emails and maintain his tracking spreadsheet [14]. In these exchanges, the automated participant was the one that classified the request correctly.
McDonald's is the other kind of miss. It is the one product teams should recognize. The 515-page file described his app interactions in granular detail and included a prediction that he would never stop eating there [1]. That is a retention model delivered as a document dump. It is responsive to the statute and close to unusable for the person holding it. Ben Winters, director of AI and privacy at the Consumer Federation of America, told WIRED the handling of these requests was "crazy" and "not an acceptable status quo" [12].
Two axes are worth drawing over your own request flow: whether intake distinguishes access from deletion, and whether the first action an agent is able to take is reversible. Suppressing a listing from search results [10] sits in the recoverable half. Deleting an account the requester then has to register for again [7] does not. If your classification is weak, the cheap protection is making the default action reversible; if your defaults are destructive, the classification has to be right on the first read. Learning which cell you occupy costs an afternoon: someone outside the privacy function files a request from a personal address, and you look at what the first action taken was and whether anyone could undo it.
